• Ubuntu 24.04
• Python 3.8

The point cloud is recursively partitioned using binary clustering.

Both standard KMeans and balanced KMeans were tested. The final implementation uses standard KMeans.

• Input mesh: welsh-dragon-small-centered.stl
• Number of levels: 10
• Output: binary_level_{level}.vtk (10 files total)

Each level’s binary labels are stored in the "clusters" attribute of the .vtk file.

python generate_training_data.py \
    --output_root ./dataset_normal \
    --start_id 0 \
    --num_samples 2000

This step:

• samples 2000 random seeds
• generates random camera 6DoF poses
• applies rotation and translation to the mesh
• renders X-ray images
• produces corresponding masks and 10-bit binary codes

dataset/
├── sample_000000/
│   ├── xray.png
│   ├── mask.png
│   ├── mask.npy
│   ├── code_stack.npy
│   ├── packed_code.npy
│   ├── code_vis.png
│   └── sample_meta.json
├── sample_000001/
│   ├── xray.png
│   ├── mask.png
│   ├── mask.npy
│   ├── code_stack.npy
│   ├── packed_code.npy
│   ├── code_vis.png
│   └── sample_meta.json
...

• xray.png
  Input image to the neural network.

• mask.npy (H, W)

  • 0 = background
  • 1 = object

• code_stack.npy (10, H, W)

  • contains 10 binary code channels
  • only valid where mask == 1
  • background is all zeros

• packed_code.npy
  Encoded as:

  code = b0 + 2b1 + 4b2 + ... + 2^9 b9
  

• code_vis.png
  Visualization for quick sanity checking.

python train.py \
    --data_root ./dataset_test \
    --save_dir ./checkpoints \
    --epochs 50 \
    --batch_size 8 \
    --lr 1e-3 \
    --backbone resnet34

python train.py \
    --data_root ./dataset_normal \
    --save_dir ./checkpoints \
    --epochs 50 \
    --batch_size 8 \
    --lr 1e-3 \
    --backbone resnet34 \
    --alpha 1.0 \
    --sigma 3.0 \
    --momentum 0.9 \
    --use_pred_mask_for_code

python train.py \
    --data_root ./dataset_normal \
    --save_dir ./checkpoints \
    --epochs 50 \
    --batch_size 8 \
    --lr 3e-4 \
    --backbone resnet34 \
    --alpha 1.5 \
    --sigma 5.0 \
    --momentum 0.7 \
    --use_pred_mask_for_code

python train_resnet_codebook.py \
    --data_root ./dataset \
    --save_dir ./checkpoints \
    --epochs 50 \
    --batch_size 8 \
    --lr 1e-3 \
    --backbone resnet18 \
    --pretrained

After training, save_dir will contain:

• history.json / history.csv
• best_metrics.json — best epoch and corresponding metrics
• last.pth / best.pth — model checkpoints

python analyze_training_results.py \
    --save_dir ./checkpoints \
    --out_dir ./checkpoints/analysis

• Binary codes are only meaningful inside object regions (mask == 1).
• Background pixels are always zero.
• code_vis.png is intended only for visualization, not training.